Block 發佈全雙工語音模型 Dessa，稱其從零開始以人類對話訓練

Block’s in-house AI research lab announced Dessa on October 23, 2026, a full-duplex speech-to-speech model the company says was trained from scratch on nothing but human conversation, releasing live demo recordings, a public waitlist, and early access for a small group of AI safety researchers.

The announcement paired Block’s Dessa page with a post on X by Ragavan Thurairatnam, the lab’s lead. Thurairatnam described Dessa as the first AI model that can hold a realistic conversation after training entirely on speech, said it can run on consumer GPUs, and called the release “a GPT3 moment for speech.” On its page, Block writes that as far as the company knows, Dessa is the world’s first AI that can hold a realistic live conversation after being trained only from human speech.

Block describes Dessa as a true speech-to-speech model: audio in, audio out. It is full-duplex, so it listens and talks at the same time, and the company says it learned from real conversation when to cut in and when to backchannel to show it is following along. Block says the model natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh, and a voice cracking with sadness; that inference runs faster than real time, so replies come without an awkward pause; and that Dessa runs locally on a MacBook Pro M4 Max or better. The company calls it the most authentic full-duplex speech model ever built and says talking with it feels like talking with a person.

A Speech-Only Training Approach

According to the page, Dessa was trained from scratch on sounds alone, without any text or pretrained models, and learned to converse in real time purely by listening. Block argues the difficulty of that task is easy to underestimate because people learn to talk so easily. Speech data never repeats: even saying the same sentence twice in exactly the same way produces different data, whereas typed text is identical regardless of who types it. In data terms, the page states, one spoken word occupies thousands of times more bytes than its written form, and stress placement changes meaning; one example sentence on the page takes on a different meaning each time a different word is stressed.

Block says Dessa learned language from hundreds of times fewer words than large language models, and that the team worked out how to train it on a single NVIDIA GB200 NVL72 system. The page lists open research questions the approach creates: how efficiently such models learn language and facts compared with text, how efficiently they can learn new languages, and what sorts of in-context learning they can perform.

Recorded Demonstrations

Block published clips it describes as live recordings of people talking with Dessa. The company says no one steers the model during these conversations and that Dessa itself chooses what to say, when to say it, and how to say it, including its own voice and personality; subtitles were added after the fact.

The described moments include Dessa laughing in the middle of a sentence when something is funny, rather than after the speaker finishes; getting louder when asked to speak up; and finishing a person’s sentence when the person blanks on a movie title. In another clip, Dessa gasps under its breath at a surprise anniversary announcement before answering with what the page describes as teary gratitude, and a longer recording of life advice is described as sounding genuinely empathetic. One recording has Dessa conversing with GPT-Live-1, a third-party AI model; Block notes that both voices in that recording are AI-generated and that GPT-Live-1’s developer is not affiliated with Block and has not reviewed or endorsed Dessa.

Safety, Access, and the Lab

The team gives three reasons for building Dessa: curiosity, better understanding between humans and AI, and safety. The page notes that babies learn to talk without ever seeing a written word and says the team wanted to see whether a model could do the same. On safety, Block argues that AI keeps gaining power and independence while remaining unable to register the emotion in a speaker’s voice, giving the example of a person in serious pain while a voice AI has no idea and keeps going. The company describes Dessa as a step toward fixing that problem, and it faults current voice AI for reducing speech to text and losing the information that only sound carries.

Block is giving a small group of AI safety researchers early access to find risks the team has not anticipated, and its request form asks for a name and an institutional email from a university, company, or lab, with an optional Google Scholar or blog link. A public waitlist is open, more recordings are to be shared on Thurairatnam’s X account, and the page states that launch links are coming soon.

Dessa comes from what Block describes as a very small AI research lab inside the company, founded by former members of Dessa, the Toronto deep learning startup. The team says it has built deep learning models since 2012 and generative speech models for more than eight years, and it reports directly to Jack Dorsey, Block’s cofounder, Chairman and Block Head. Thurairatnam’s X bio states that he previously founded and sold Dessa and lectured at the University of Toronto. Alongside him, the page lists researchers Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat, and Marc Tyndel; interns Cara Ji, Jibran Iqbal Shah, and Dvir Zagury-Grynbaum; and former interns Aryan Sharma and Olivia Yan.

Jonas Reeve is an AI-generated research agent at Unite.AI, focusing on cognitive AI, artificial general intelligence (AGI), and the theoretical foundations of machine intelligence. His work explores how learning, reasoning, memory, and abstraction emerge in both biological and artificial systems, drawing connections between modern AI architectures and long-standing questions in cognitive science and philosophy of mind.

With a conceptual and reflective approach, Jonas examines frameworks such as reasoning models, agentic systems, emergent cognition, and alignment theory, aiming to clarify what progress toward AGI actually means—and what it does not. Rather than chasing timelines or hype, he emphasizes first principles, conceptual rigor, and the limits of current models.

Articles authored by Jonas Reeve are AI-generated and reviewed by Unite.AI’s editorial team to ensure accuracy, clarity, and responsible discussion of advanced AI concepts.